Text abstract generation method and device, equipment and storage medium
By using unsupervised learning methods to segment and encode articles, generating input vectors and performing data augmentation, and combining multilayer perceptron and CTC beam search, the problem of low efficiency in text summarization generation is solved, achieving efficient and accurate summarization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- PING AN TECH (SHENZHEN) CO LTD
- Filing Date
- 2022-07-28
- Publication Date
- 2026-06-02
AI Technical Summary
Existing text summarization techniques are inefficient, especially generative methods that rely on labeled data to train models, resulting in summaries that do not match the article's topic and are inefficient.
The article is segmented and encoded using an unsupervised learning method to generate an input vector. Three types of augmented data are obtained through data augmentation. The maximum probability path is generated by combining a multilayer perceptron and CTC's bundle search to form a summary.
It achieves improved efficiency and accuracy of text summarization without the need for manual data annotation, realizing automated text generation and efficient summary generation.
Smart Images
Figure CN115203374B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a text summarization method, apparatus, electronic device, and computer-readable storage medium. Background Technology
[0002] For purposes such as e-book retrieval and information browsing, summarization technology is a method of generating summaries from article data to achieve these goals. Existing text summarization technologies mainly fall into two categories: one is the extractive method, which primarily calculates the importance of sentences in an article and extracts the most important sentences as summaries. A drawback of this method is that it often results in incomplete sentence extractions. The other is the generative method, which inputs all the words in the article into an encoder model for encoding, and then uses a decoder to generate sentence summaries. The second method suffers from a strong correlation between the quality of the decoder and the training of the language model. Poorly trained models often result in generated sentences that do not match the article's topic. Furthermore, model training requires learning parameters using labeled data, which necessitates manual data annotation, leading to low efficiency in model-based text summarization. Summary of the Invention
[0003] This invention provides a text summarization method, apparatus, electronic device, and computer-readable storage medium, the main purpose of which is to solve the problem of low efficiency in text summarization generation.
[0004] To achieve the above objectives, the present invention provides a text summarization method, comprising:
[0005] The acquired article to be processed is segmented into words. A dictionary is constructed based on the results of the word segmentation. The words in the dictionary are encoded to obtain the feature vectors corresponding to the words. The feature vectors are then used as input vectors.
[0006] The input vector is augmented to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector.
[0007] Network computation is performed on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector;
[0008] When the number of data augmentations reaches a preset number, the update vector is associated with the dictionary, and vector calculation is performed on the associated dictionary to obtain the probability vector corresponding to the update vector.
[0009] A global probability path is constructed based on the probability vector, and the maximum probability path is obtained based on the global probability path to form a summary of the article to be processed.
[0010] Optionally, the step of performing word segmentation on the acquired article to be processed includes:
[0011] Target sentences are selected one by one from the article to be processed, and word segmentation is performed on the target sentences to obtain the word segments of multiple first words;
[0012] Based on the part-of-speech tag of the first segment, the target sentence is syntactically analyzed to obtain multiple second segments, and stop words in the second segments are deleted;
[0013] The second word segment is filtered and re-segmented using a pre-built custom dictionary to obtain multiple word segments of the article to be processed.
[0014] Optionally, the step of performing data augmentation on the input vector to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector includes:
[0015] Each input vector is copied three times to obtain three copied vectors corresponding to the input vector;
[0016] Each of the copied vectors is calculated with a randomly generated initialization matrix to obtain the first augmented data, the second augmented data, and the third augmented data of the input vector.
[0017] Optionally, the step of performing network computation on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector includes:
[0018] Target input vectors are selected one by one from the input vectors, and network calculations are performed using the first augmented data of the target input vectors and the second augmented data of each input vector to obtain the attention weight value of each input vector corresponding to the target input vector;
[0019] The attention weight value is multiplied by the third augmentation data of each corresponding input vector to obtain the third augmentation update data of each input vector corresponding to the target input vector;
[0020] The third enhanced update data of each input vector is summed to obtain the update vector of the target input vector.
[0021] Optionally, the step of performing network computation using the first augmented data of the target input vector and the second augmented data of each input vector to obtain the attention weight value of each input vector corresponding to the target input vector includes:
[0022] The first augmented data is multiplied by the second augmented data of each input vector to obtain the inner product value.
[0023] The inner product value is regularized, and the regularized inner product value is input into a normalization function to obtain the attention weight value of each input vector.
[0024] Optionally, the step of performing vector calculation on the associated dictionary to obtain the probability vector corresponding to the update vector includes:
[0025] The input layer of a pre-defined multilayer perceptron is constructed based on the update vector corresponding to each word in the dictionary and special characters.
[0026] The input layer is convolved using the hidden layers and fully connected layers of the multilayer perceptron, and the calculation result is output through the softmax function to obtain the output vector;
[0027] The length of the output vector is set according to the length of the dictionary to obtain the probability vector corresponding to each word segmentation, wherein the value of the probability vector represents the probability value of the word segmentation.
[0028] Optionally, obtaining the maximum probability path based on the global probability path to form a summary of the article to be processed includes:
[0029] The first-level bundle nodes in the global probability path are used as the basic nodes;
[0030] Determine whether there are null values in the vector values of the next level node of the basic node;
[0031] If there is a null value, the vector value of the base node is merged with the vector value of the next level node to obtain the updated vector value;
[0032] If no null value exists, it will be determined whether the vector value of the next-level node is the same as the vector value of the base node.
[0033] If they are the same, the vector value of the basic node is merged with the vector value of the next level node to obtain the updated vector value;
[0034] If they are not the same, a target vector value will be selected from the vector values of the next-level node, and the target vector value will be multiplied by the vector value of the base node. The result of the multiplication will be used as the update vector value of the next-level node.
[0035] Select an update vector value that meets the preset selection criteria from the updated vector values and use it as the new vector value for the next level node;
[0036] The next-level node is used as the new base node, and the step of determining whether there is a null value in the vector value of the next-level node of the base node is returned, until all the bundle nodes of the global probability path have been traversed.
[0037] The new vector value in the next level node corresponding to the last basic node is output as a summary.
[0038] To address the above problems, the present invention also provides a text summarization generation apparatus, the apparatus comprising:
[0039] The input vector generation module is used to perform word segmentation on the acquired article to be processed, construct a dictionary based on the word segmentation results, encode the words in the dictionary to obtain the feature vectors corresponding to the words, and use the feature vectors as input vectors.
[0040] The data augmentation module is used to augment the input vector to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector.
[0041] The update vector generation module is used to perform network calculations on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector;
[0042] The probability vector generation module is used to associate the update vector with the dictionary when the number of data augmentations reaches a preset number, and to perform vector calculation on the associated dictionary to obtain the probability vector corresponding to the update vector.
[0043] The abstract generation module is used to construct a global probability path based on the probability vector, and obtain the maximum probability path based on the global probability path to form an abstract of the article to be processed.
[0044] To address the above problems, the present invention also provides an electronic device, the electronic device comprising:
[0045] At least one processor; and,
[0046] A memory communicatively connected to the at least one processor; wherein,
[0047] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the text summarization generation method described above.
[0048] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one computer program, which is executed by a processor in an electronic device to implement the text summarization generation method described above.
[0049] This invention employs an unsupervised learning method, enabling the model to generate text summaries without manual data annotation, thus improving the efficiency of text summarization. By segmenting and encoding the text, an input vector is generated for the model. Data augmentation is applied to the input vector, resulting in three types of augmented data. These three augmented data are then combined for comprehensive calculation to obtain an update vector, which is then iteratively generated to produce new update vectors. This avoids manual data annotation, improves model computation efficiency, and automates text generation. Furthermore, by using a multilayer perceptron and CTC beam search, the maximum probability path is generated, enabling the generation of summaries from the text and improving the accuracy of the generated summaries. Therefore, the text summarization method, apparatus, electronic device, and computer-readable storage medium proposed in this invention can solve the problem of low efficiency in text summarization. Attached Figure Description
[0050] Figure 1 This is a flowchart illustrating a text summarization method according to an embodiment of the present invention.
[0051] Figure 2 This is a schematic diagram of the process of performing network calculations on first enhanced data, second enhanced data and third enhanced data according to an embodiment of the present invention;
[0052] Figure 3 This is a schematic diagram illustrating the process of performing vector calculations on the associated dictionary according to an embodiment of the present invention;
[0053] Figure 4 A functional block diagram of a text summarization generation device provided in an embodiment of the present invention;
[0054] Figure 5 This is a schematic diagram of the structure of an electronic device that implements the text summarization method according to an embodiment of the present invention.
[0055] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0056] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0057] This application provides a text digest generation method. The execution entity of the text digest generation method includes, but is not limited to, at least one of the following: a server, a terminal, or an electronic device that can be configured to execute the method provided in this application. In other words, the text digest generation method can be executed by software or hardware installed on a terminal device or a server device, and the software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.
[0058] Reference Figure 1 The diagram shown is a flowchart illustrating a text summarization generation method according to an embodiment of the present invention. In this embodiment, the text summarization generation method includes:
[0059] S1. Perform word segmentation on the acquired article to be processed, construct a dictionary based on the word segmentation results, encode the words in the dictionary to obtain the feature vectors corresponding to the words, and use the feature vectors as input vectors.
[0060] In this embodiment of the invention, the article to be processed can be a thesis, news article, magazine article, etc.
[0061] In this embodiment of the invention, the step of performing word segmentation on the acquired article to be processed includes:
[0062] Target sentences are selected one by one from the article to be processed, and word segmentation is performed on the target sentences to obtain the word segments of multiple first words;
[0063] Based on the part-of-speech tag of the first segment, the target sentence is syntactically analyzed to obtain multiple second segments, and stop words in the second segments are deleted;
[0064] The second word segment is filtered and re-segmented using a pre-built custom dictionary to obtain multiple word segments of the article to be processed.
[0065] Specifically, in this embodiment of the invention, the hanlp word segmentation tool can be used to perform part-of-speech segmentation on the target sentence; the custom dictionary can be existing or popular words crawled from the Internet using big data technology or web crawling technology.
[0066] For example, suppose the target sentence is "Machine learning drives the rapid development of the artificial intelligence industry." Part-of-speech analysis of the target sentence yields "Machine (n.) learning (v.) drives (v.) artificial intelligence (n.) industry (n.) rapid development (nz.) (w.)", where (n.), (v.), (nz.), and (w.) represent the parts of speech tagged by the hanlp word segmentation tool, and "machine", "learning", "drive", "artificial intelligence", "industry", "rapid development", and "." are the first segmented words. Further, syntactic analysis is performed based on the parts of speech of the first segmented words to obtain the first part of speech of the target sentence. The word segmentation contains two predicate verbs, "learn" and "drive," where "machine" and "learn" have a subject-predicate relationship, and "learn" and "drive" are parallel predicates. Therefore, "machine" and "learn" can be merged into "machine learning" as the second word segment, or "learn" and "drive" can be merged into "learning drives" as the second word segment, and the stop word "." should be deleted. The word "machine learning" exists in the custom dictionary. Therefore, by filtering the second word segmentation through the custom dictionary, the following multiple word segments are obtained for the article to be processed: "machine learning," "drive," "artificial intelligence," "industry," and "rapid development."
[0067] In this embodiment of the invention, the word segments are added to a preset dictionary, and a preset special dictionary is randomly added to the dictionary. The dictionary is composed of the word segments and the special characters. The special characters can be empty words, and the dictionary length can be represented by the number of word segments.
[0068] In this embodiment of the invention, the dictionary can be segmented using the glove language model, and the specific process may include vectorization, vector enhancement, etc.
[0069] In this embodiment of the invention, encoding the word segments in the dictionary to obtain the feature vectors corresponding to the word segments includes:
[0070] The word segmentation is subjected to grammatical analysis to generate the identifier sequence corresponding to the dictionary;
[0071] The word2vec network is used to vectorize and extract convolutional features from each identifier in the identifier sequence to obtain the feature vector corresponding to each word segmentation.
[0072] In this embodiment of the invention, a custom grammar analyzer can be used to perform grammar analysis on the word segmentation; the identifiers in the identifier sequence can be classified as: string values, single character values, character arrays, numbers, operators, names, etc.
[0073] In another optional embodiment of the present invention, a bag-of-words model can be used to perform vector processing on each identifier in the identifier sequence to obtain the feature vector corresponding to each word segmentation.
[0074] In this embodiment of the invention, the word2vec network embedding is used to embed low-dimensional information into a high-dimensional space, thereby realizing vector computation in the network and improving the vector representation of information through high-dimensional space mapping.
[0075] S2. Perform data augmentation on the input vector to obtain the first augmented data, the second augmented data, and the third augmented data corresponding to the input vector.
[0076] In this embodiment of the invention, data augmentation methods may include vector copying, cross-computation, etc.
[0077] In this embodiment of the invention, the step of performing data augmentation on the input vector to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector includes:
[0078] Each input vector is copied three times to obtain three copied vectors corresponding to the input vector;
[0079] Each of the copied vectors is calculated with a randomly generated initialization matrix to obtain the first augmented data, the second augmented data, and the third augmented data of the input vector.
[0080] For example, given input vectors A, B, and C, taking input vector A as an example, input vector A is copied three times to obtain copied vectors A1, A2, and A3. A 512×512 initialization matrix is randomly generated. This initialization matrix is then multiplied by copied vectors A1, A2, and A3 to obtain the first augmented data Q. a Second augmented data K a And the third enhanced data V a .
[0081] S3. Perform network computation on the first enhanced data, the second enhanced data, and the third enhanced data to obtain the update vector of the input vector.
[0082] In this embodiment of the invention, the first output vector can be used as the first target output vector, and the first augmented data corresponding to the target output vector is the standard augmented data. It can be assumed that the input vector A is the first output vector, and its corresponding first augmented data Q... a Enhanced Data Q for Standards θ .
[0083] Please see Figure 2As shown in this embodiment of the invention, the step of performing network computation on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector includes:
[0084] S21. Select target input vectors one by one from the input vectors, and perform network calculations using the first augmented data of the target input vector and the second augmented data of each input vector to obtain the attention weight value of each input vector corresponding to the target input vector;
[0085] S22. Multiply the attention weight value with the third augmentation data of each corresponding input vector to obtain the third augmentation update data of each input vector corresponding to the target input vector;
[0086] S23. Add the third enhanced update data of each input vector to obtain the update vector of the target input vector.
[0087] Further, the step of performing network computation using the first augmented data of the target input vector and the second augmented data of each input vector to obtain the attention weight value of each input vector corresponding to the target input vector includes:
[0088] The first augmented data is multiplied by the second augmented data of each input vector to obtain the inner product value.
[0089] The inner product value is regularized, and the regularized inner product value is input into a normalization function to obtain the attention weight value of each input vector.
[0090] In this embodiment of the invention, the normalization function includes, but is not limited to, the softmax function.
[0091] For example, there exist input vectors A, B, and C, and the target input vector is assumed to be input vector A, which corresponds to the first augmentation data Q. a Second augmented data K a And the third enhanced data V a The input vector B corresponds to the first augmented data Q. b Second augmented data K b And the third enhanced data V b The input vector C corresponds to the first augmented data Q. c Second augmented data K c And the third enhanced data V c Using the first augmented data Q of the target input vector A. a The second augmented data K is used in conjunction with input vectors A, B, and C. a K b Kc Perform network computation to obtain the attention weight values W corresponding to input vectors A, B, and C. a W b W c The attention weight value W a W b W c The third augmented data V, along with input vectors A, B, and C. a V b V c Multiplying these results in the third augmented update data V, which is the sum of input vectors A, B, and C. + a V + b V + c The third augmentation update data V is applied to input vectors A, B, and C. + a V + b V + c The two vectors are added together to obtain the update vector V of the target input vector A. ' a By successively using input vectors B and C as target input vectors, the update vector V of input vectors B and C can be obtained. ' b and V ' c .
[0092] In this embodiment of the invention, the number of times data augmentation is performed can be set according to model calculation requirements or vector filtering requirements, for example, the number of times data augmentation is performed can be two.
[0093] In this embodiment of the invention, after obtaining the updated vector of the input vector, the method further includes: determining whether the data augmentation has reached a preset number of times. Further, if the number of data augmentations has not reached the preset number, the updated vector is used as a new input vector, and the process returns to S2 above; if the number of data augmentations has reached the preset number, the updated vector is processed.
[0094] In this embodiment of the invention, if the data augmentation has not reached the preset number of times, data augmentation needs to be performed again, that is, the steps S2 to S3 above are returned to perform data augmentation again to obtain the required update vector.
[0095] S4. When the number of data augmentations reaches a preset number, the update vector is associated with the dictionary, and vector calculation is performed on the associated dictionary to obtain the probability vector corresponding to the update vector.
[0096] In this embodiment of the invention, it is assumed that the data augmentation can be performed twice. When the data augmentation reaches a preset number of times, that is, when two data augmentations are completed, the updated vector V of the input vectors A, B, and C is obtained. '' a V '' b and V '' c .
[0097] In this embodiment of the invention, the update vector can be associated with the word segmentation in the dictionary, thereby achieving association with the dictionary.
[0098] Please see Figure 3 As shown in this embodiment of the invention, the step of performing vector calculation on the associated dictionary to obtain the probability vector corresponding to the update vector includes:
[0099] S31. Construct a preset input layer of a multilayer perceptron based on the update vector corresponding to each word in the dictionary and special characters;
[0100] S32. The input layer is convolved using the hidden layer and fully connected layer of the multilayer perceptron, and the calculation result is output through the softmax function to obtain the output vector;
[0101] S33. Set the length of the output vector according to the length of the dictionary to obtain the probability vector corresponding to each word segmentation, wherein the value of the probability vector represents the probability value of the word segmentation.
[0102] In this embodiment of the invention, the length of the dictionary is equal to the number of words in the dictionary; the sum of the values of the probability vectors is 1.
[0103] S5. Construct a global probability path based on the probability vector, and obtain the maximum probability path based on the global probability path to form a summary of the article to be processed.
[0104] In this embodiment of the invention, constructing a global probability path based on the probability vector includes:
[0105] Select target probability vectors one by one from the probability vectors, and use the selected target probability vectors as multi-level bundle nodes in sequence;
[0106] A preset number of vector values are selected from the target probability vector corresponding to each bundle node;
[0107] A global probability path is constructed based on the bundle nodes and the selected vector values.
[0108] In this embodiment of the invention, the first target probability vector can be used as a first-level bundle node, and the subsequently selected target probability vectors can be used as second-level bundle nodes, third-level bundle nodes, etc.; in this embodiment of the invention, the three vector values with the largest vector values in the target probability vector can be selected.
[0109] In this embodiment of the invention, obtaining the maximum probability path based on the global probability path to form a summary of the article to be processed includes:
[0110] The first-level bundle nodes in the global probability path are used as the basic nodes;
[0111] Determine whether there are null values in the vector values of the next level node of the basic node;
[0112] If there is a null value, the vector value of the base node is merged with the vector value of the next level node to obtain the updated vector value;
[0113] If no null value exists, it will be determined whether the vector value of the next-level node is the same as the vector value of the base node.
[0114] If they are the same, the vector value of the basic node is merged with the vector value of the next level node to obtain the updated vector value;
[0115] If they are not the same, a target vector value will be selected from the vector values of the next-level node, and the target vector value will be multiplied by the vector value of the base node. The result of the multiplication will be used as the update vector value of the next-level node.
[0116] Select an update vector value that meets the preset selection criteria from the updated vector values and use it as the new vector value for the next level node;
[0117] The next-level node is used as the new base node, and the step of determining whether there is a null value in the vector value of the next-level node of the base node is returned, until all the bundle nodes of the global probability path have been traversed.
[0118] The new vector value in the next level node corresponding to the last basic node is output as a summary.
[0119] In this embodiment of the invention, the selection criteria can be the three values with the largest update vector values.
[0120] In this embodiment of the invention, the beam search method in CTC can be used to search for the path with the highest global probability among all probability vectors, which can then be used as the summary of the article to be processed.
[0121] This invention employs an unsupervised learning method, enabling the model to generate text summaries without requiring manually labeled data, thus improving the efficiency of text summarization. It generates the model's input vector by segmenting and encoding the text into words. Data augmentation is then applied to the input vector, resulting in three types of augmented data. These three augmented data are combined for comprehensive calculation to obtain an update vector, which is then iteratively generated to produce new update vectors. This avoids manual data labeling, improves model computation efficiency, and automates text generation. Furthermore, by using a multilayer perceptron and CTC's bundle search, the maximum probability path is generated, enabling the generation of summaries from the text and improving the accuracy of the generated summaries. Therefore, the text summarization method proposed in this invention can solve the problem of low efficiency in text summarization.
[0122] like Figure 4 The diagram shown is a functional block diagram of a text summarization generation device provided in an embodiment of the present invention.
[0123] The text summarization generation device 100 of this invention can be installed in an electronic device. Depending on the functions implemented, the text summarization generation device 100 may include an input vector generation module 101, a data augmentation module 102, an update vector generation module 103, a probability vector generation module 104, and a summarization generation module 105. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and which are stored in the memory of the electronic device.
[0124] In this embodiment, the functions of each module / unit are as follows:
[0125] The input vector generation module 101 is used to perform word segmentation on the acquired article to be processed, construct a dictionary based on the result of word segmentation, encode the words in the dictionary, obtain the feature vectors corresponding to the words, and use the feature vectors as input vectors.
[0126] The data augmentation module 102 is used to perform data augmentation on the input vector to obtain first augmented data, second augmented data and third augmented data corresponding to the input vector;
[0127] The update vector generation module 103 is used to perform network calculations on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector;
[0128] The probability vector generation module 104 is used to associate the update vector with the dictionary when the number of data augmentations reaches a preset number, and to perform vector calculation on the associated dictionary to obtain the probability vector corresponding to the update vector.
[0129] The summary generation module 105 is used to construct a global probability path based on the probability vector, and obtain the maximum probability path based on the global probability path to form a summary of the article to be processed.
[0130] In detail, each module in the text summarization generation device 100 described in this embodiment of the invention uses the same technical means as the text summarization generation method described in the accompanying drawings and can produce the same technical effect, which will not be repeated here.
[0131] like Figure 5 The diagram shown is a structural schematic of an electronic device for implementing a text summarization method according to an embodiment of the present invention.
[0132] The electronic device 1 may include a processor 10, a memory 11, a communication bus 12 and a communication interface 13, and may also include a computer program, such as a text summarization program, stored in the memory 11 and capable of running on the processor 10.
[0133] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., executing text digest generation programs) and calls data stored in the memory 11 to perform various functions of the electronic device and process data.
[0134] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as the code of a text summarization program, but also to temporarily store data that has been output or will be output.
[0135] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.
[0136] The communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), or optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.
[0137] Figure 5 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 5The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.
[0138] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.
[0139] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.
[0140] The text digest generation program stored in the memory 11 of the electronic device 1 is a combination of multiple instructions, which, when run in the processor 10, can achieve the following:
[0141] The acquired article to be processed is segmented into words. A dictionary is constructed based on the results of the word segmentation. The words in the dictionary are encoded to obtain the feature vectors corresponding to the words. The feature vectors are then used as input vectors.
[0142] The input vector is augmented to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector.
[0143] Network computation is performed on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector;
[0144] When the number of data augmentations reaches a preset number, the update vector is associated with the dictionary, and vector calculation is performed on the associated dictionary to obtain the probability vector corresponding to the update vector.
[0145] A global probability path is constructed based on the probability vector, and the maximum probability path is obtained based on the global probability path to form a summary of the article to be processed.
[0146] Specifically, the specific implementation method of the processor 10 for the above instructions can be referred to the description of the relevant steps in the corresponding embodiment of the accompanying drawings, and will not be repeated here.
[0147] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).
[0148] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following:
[0149] The acquired article to be processed is segmented into words. A dictionary is constructed based on the results of the word segmentation. The words in the dictionary are encoded to obtain the feature vectors corresponding to the words. The feature vectors are then used as input vectors.
[0150] The input vector is augmented to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector.
[0151] Network computation is performed on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector;
[0152] When the number of data augmentations reaches a preset number, the update vector is associated with the dictionary, and vector calculation is performed on the associated dictionary to obtain the probability vector corresponding to the update vector.
[0153] A global probability path is constructed based on the probability vector, and the maximum probability path is obtained based on the global probability path to form a summary of the article to be processed.
[0154] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.
[0155] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0156] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.
[0157] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0158] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.
[0159] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0160] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.
[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A text summarization method, characterized in that, The method includes: The acquired article to be processed is segmented into words. A dictionary is constructed based on the results of the word segmentation. The words in the dictionary are encoded to obtain the feature vectors corresponding to the words. The feature vectors are then used as input vectors. The input vector is augmented to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector. Network computation is performed on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector; When the number of data augmentations reaches a preset number, the update vector is associated with the dictionary, and vector calculation is performed on the associated dictionary to obtain the probability vector corresponding to the update vector. A global probability path is constructed based on the probability vector, and the maximum probability path is obtained based on the global probability path to form a summary of the article to be processed. The step of performing network computation on the first enhanced data, the second enhanced data, and the third enhanced data to obtain the update vector of the input vector includes: Target input vectors are selected one by one from the input vectors, and network calculations are performed using the first augmented data of the target input vectors and the second augmented data of each input vector to obtain the attention weight value of each input vector corresponding to the target input vector; The attention weight value is multiplied by the third augmentation data of each corresponding input vector to obtain the third augmentation update data of each input vector corresponding to the target input vector; The third enhanced update data of each input vector is summed to obtain the update vector of the target input vector.
2. The text summarization method as described in claim 1, characterized in that, The word segmentation process for the acquired article to be processed includes: Target sentences are selected one by one from the article to be processed, and word segmentation is performed on the target sentences to obtain the word segments of multiple first words; Based on the part-of-speech tag of the first segment, the target sentence is syntactically analyzed to obtain multiple second segments, and stop words in the second segments are deleted; The second word segment is filtered and re-segmented using a pre-built custom dictionary to obtain multiple word segments of the article to be processed.
3. The text summarization method as described in claim 1, characterized in that, The step of performing data augmentation on the input vector to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector includes: Each input vector is copied three times to obtain three copied vectors corresponding to the input vector; Each of the copied vectors is calculated with a randomly generated initialization matrix to obtain the first augmented data, the second augmented data, and the third augmented data of the input vector.
4. The text summarization method as described in claim 1, characterized in that, The step of performing network computation using the first augmented data of the target input vector and the second augmented data of each input vector to obtain the attention weight value of each input vector corresponding to the target input vector includes: The first augmented data is multiplied by the second augmented data of each input vector to obtain the inner product value. The inner product value is regularized, and the regularized inner product value is input into a normalization function to obtain the attention weight value of each input vector.
5. The text summarization method as described in claim 1, characterized in that, The step of performing vector calculations on the associated dictionary to obtain the probability vector corresponding to the update vector includes: The input layer of a pre-defined multilayer perceptron is constructed based on the update vector corresponding to each word in the dictionary and special characters. The input layer is convolved using the hidden layers and fully connected layers of the multilayer perceptron, and the calculation result is output through the softmax function to obtain the output vector; The length of the output vector is set according to the length of the dictionary to obtain the probability vector corresponding to each word segmentation, wherein the value of the probability vector represents the probability value of the word segmentation.
6. The text summarization method according to any one of claims 1 to 5, characterized in that, The step of obtaining the maximum probability path based on the global probability path to form a summary of the article to be processed includes: The first-level bundle nodes in the global probability path are used as the basic nodes; Determine whether there are null values in the vector values of the next level node of the basic node; If there is a null value, the vector value of the base node is merged with the vector value of the next level node to obtain the updated vector value; If no null value exists, it will be determined whether the vector value of the next-level node is the same as the vector value of the base node. If they are the same, the vector value of the basic node is merged with the vector value of the next level node to obtain the updated vector value; If they are not the same, a target vector value will be selected from the vector values of the next-level node, and the target vector value will be multiplied by the vector value of the base node. The result of the multiplication will be used as the update vector value of the next-level node. Select an update vector value that meets the preset selection criteria from the updated vector values and use it as the new vector value for the next level node; The next-level node is used as the new base node, and the step of determining whether there is a null value in the vector value of the next-level node of the base node is returned, until all the bundle nodes of the global probability path have been traversed. The new vector value in the next level node corresponding to the last basic node is output as a summary.
7. A text summarization generation apparatus for implementing the text summarization generation method as described in any one of claims 1 to 6, characterized in that, The device includes: The input vector generation module is used to perform word segmentation on the acquired article to be processed, construct a dictionary based on the word segmentation results, encode the words in the dictionary to obtain the feature vectors corresponding to the words, and use the feature vectors as input vectors. The data augmentation module is used to augment the input vector to obtain first augmented data, second augmented data, and third augmented data corresponding to the input vector. The update vector generation module is used to perform network calculations on the first augmented data, the second augmented data, and the third augmented data to obtain the update vector of the input vector; The probability vector generation module is used to associate the update vector with the dictionary when the number of data augmentations reaches a preset number, and to perform vector calculation on the associated dictionary to obtain the probability vector corresponding to the update vector. The abstract generation module is used to construct a global probability path based on the probability vector, and obtain the maximum probability path based on the global probability path to form an abstract of the article to be processed.
8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the text summarization generation method as described in any one of claims 1 to 6.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the text summarization method as described in any one of claims 1 to 6.